Commercial lease agreements can be tedious to analyse, with their extensive length (40-60 pages) and obscure language. Our AI solution simplifies and automates the preparatory tasks involved in sifting through them. Users can effortlessly send large volumes of documents to an email address and receive the extracted information in the desired format. This AI service integrates external services like Google Sheets to allow users to customise their queries and toggle parameters. Overall, our AI solution has dramatically streamlined our client's administrative tasks, enabling them to refocus their time and energy on their legal tasks instead.

Pubs Advisory Service specialises in providing expert guidance, information, and representation services to a diverse range of clients within the pub sector. With a solid reputation as the lead advisory and guidance organisation for pubs in the UK, Pubs Advisory Service proudly holds a prominent market position and is the go-to choice for pubs seeking comprehensive legal support.

People entering the Licensed Trade sector need detailed independent advice on a wide range of issues. Access to such pre-entry advice can prove invaluable for conducting thorough due diligence and making informed decisions, ultimately setting the stage for successful pub ownership and prosperous growth. However, because each commercial lease agreement is unique and requires individual attention, this detailed advice can be costly and time-consuming to prepare.

Previous methods to perform document searches on commercial lease agreements often involved scanning them – using Adobe's inbuilt Optical Character Recognition (OCR) to transcribe handwriting – and performing traditional keyword searches (e.g. ctrl+F) throughout the PDF. However, this method is flawed on several fronts:

• Commercial lease agreements often contain handwritten notes, which are not accurately transcribed when using standard OCR software.
• The keywords used in the search bar may not appear in the text exactly as searched, and ctrl+F-type searches are not complex enough to detect variations in phrasing or understand natural language questions instead.
• The information being searched for is not usually accessible in one place, but is instead scattered throughout the entire document, requiring the user to piece it together manually.

Pubs Advisory Service realised that they needed a smarter approach to extract information from lease agreements. They asked OpenKit to develop an AI search tool able to perform natural language searches on large volumes of documents.

Consultancy

We began by conducting in-depth consultations with Pubs Advisory Service to gain a comprehensive understanding of their industry and the obstacles they encounter. OpenKit's main goal was to explore how the current landscape of generative AI could be harnessed to enhance productivity and overall business performance for Pubs Advisory Service, and provide them with a substantial competitive edge.

This unique solution required thorough exploration and prototyping. As such, a portion of the project was allocated to experimenting with various approaches, tools and methodologies.

First, OpenKit analysed the business landscape and gained a deep and nuanced understanding of the client's day-to-day business activities. We maintained an open dialogue with the client, discussing the different tools and services that we could offer, ranging from Smart Query Systems to advanced autonomous agents. Because the tool had to be user-friendly, integrating easily into the client and their customers' daily workflow, we settled on using an email interface through which users could effortlessly send questions and receive answers. We determined that avoiding the need to build a new interface from scratch would significantly reduce development time and costs for our client.

This collaborative process enabled us to establish a clear project goal and scope, which guided our research and development efforts, resulting in a refined final product.

OpenKit has existing experience using Large Language Models (LLMs) to build automation tools, so we had strong foundations to build on when approaching the problem. We quickly entered an R&D phase and identified several key aspects for the initial design:

• An email interface system capable of receiving large commercial lease agreements (up to 50mb attachments)
• A cloud-based API for Open Source and proprietary LLM inferencing, which makes the solution accessible online from any device
• State-of-the-art information extraction which could perform just as well with Open Source models (allowing clients to remain independent from third-party proprietary models)

We developed this Smart Query System iteratively, meaning we alternated between testing and development phases. This open exchange allowed us to make the system much more robust and optimise it for live use. After Pubs Advisory Service first began testing the system's performance, they provided detailed feedback. Following their report, we refined and improved the system, ensuring their maximum satisfaction.

Notable improvements included allowing users to customise their prompts without the need for any technical expertise. We did this by integrating a signposted Google Sheets interface through which they could update their questions and toggle certain parameters:

We faced a range of challenges during the completion of this project, which allowed us to demonstrate our expert problem-solving abilities.

1. OCR Limitations: The documents we initially handled were pre-scanned by Adobe's OCR service. However, the software's limitations led to multiple scanning errors, especially when it came to parsing tables and deciphering handwritten information.

   • First Solution: OpenKit's first solution was a retroactive corrective approach, using GPT-4. The LLM would parse the document a first time, highlighting OCR and formatting issues, then attempt to correct the distortions before passing the rectified output to a smaller LLM.
   • Second Solution: We introduced an AI OCR system to re-scan the full document, removing the need for Adobe's OCR. We incorporated AWS Textract into our pre-processing pipeline, which improved data quality by filling in missing information, reducing noise, and resolving inconsistencies.

2. Context Length Limitations: We faced challenges in determining the context length we needed from our chosen LLM. Most Open Source models are limited to shorter context lengths, meaning they can only consider a limited amount of text and would be incapable of handling documents as lengthy as commercial lease agreements.

   Solution: In order to make the service equally efficient on Open Source LLMs and proprietary models like GPT-4, we developed a technique which used word embeddings and separated the documents into individual 'chunks,' each capable of being processed by a more lightweight LLM.

3. Scalability: Once Pubs Advisory Service began testing the system and saw how valuable and impressive the results were, they were eager to expand its use across their customer base. However, the system was at the time incapable of serving high volumes of users simultaneously.

   Solution: OpenKit designed and developed a serverless scheduling system which queued all requests upon reception, allowing it to process them individually. This ensured a reliable, failsafe performance even when handling a dynamic amount of concurrent documents.

Our client tested our AI solution against similar existing market solutions. The results were extremely positive, as our solution outperformed competitors in terms of accuracy, reliability, and cost-effectiveness. This evaluation solidified our solution's position as the preferred choice for Pubs Advisory Service.

• Overcoming complex challenges to build a bespoke, innovative AI solution
• Outperforming existing services on key factors, including performance and price
• Establishing a standard for high scalability and reliability
• Optimising the ease of use by relying on well-known email and spreadsheet interfaces

Pubs Advisory Service was highly impressed by the results of our collaboration and has expressed an eagerness to continue working with OpenKit on several forthcoming projects.

• OpenAI LLM models
• Supabase
• Typescript
• Node.js